PharmaBench - Document Structure - Pass Rate: leaderboard
PharmaBench - Document Structure - Pass Rate: Pass rate (%). PharmaBench: 120 tasks across eight synthetic drug programs, three runs per task, same unnamed open-source coding agent in an isolated shell/Python/document sandbox without internet. The original eleven-model report uses provider-default reasoning and an unnamed LLM rubric judge; late scoreboard top-up settings are not separately reported. Pass means every required criterion met and no penalty; mean is rubric reward averaged over runs then tasks. Group breakdowns retain the eleven-model report.
Metric: Pass rate (%). Source: www.micro1.ai. Saturation forecast: Rough model projection: around 2026. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-6 Astra (PharmaBench Run: provider-default; coding-agent-unnamed) | 14.58 | |
| 2 | Claude Fable 5.1 (PharmaBench Run: provider-default; coding-agent-unnamed) | 14.58 | |
| 3 | Gemini 3.8 Flash (PharmaBench Run: provider-default; coding-agent-unnamed) | 14.58 | |
| 4 | Muse Wonder (PharmaBench Run: provider-default; coding-agent-unnamed) | 14.58 | |
| 5 | Muse Spark 1.3 (PharmaBench Run: provider-default; coding-agent-unnamed) | 12.5 | |
| 6 | Gemini 4 Argon (PharmaBench Run: provider-default; coding-agent-unnamed) | 12.5 | |
| 7 | Claude Sonnet 5.5 (PharmaBench Run: provider-default; coding-agent-unnamed) | 8.33 | |
| 8 | Claude Opus 5.5 (PharmaBench Run: provider-default; coding-agent-unnamed) | 8.33 | |
| 9 | Grok 4.6 (PharmaBench Run: provider-default; coding-agent-unnamed) | 6.25 | |
| 10 | Claude Opus 5 (PharmaBench Run: provider-default; coding-agent-unnamed) | 6.25 |
Interactive version: theaggregate.ai/benchmark?slug=pharmabench-document-structure-pass-rate · How It Works · Data refreshed daily, snapshot 2026-11-06.